SamLowe/all-MiniLM-L12-v2-go_emotions-classifier-onnx

Overview

This is a multi-label, multi-class linear classifer for emotions that works with sentence-transformers/all-MiniLM-L12-v2, having been trained on the go_emotions dataset.

Labels

The 28 labels from the go_emotions dataset are:

['admiration', 'amusement', 'anger', 'annoyance', 'approval', 'caring', 'confusion', 'curiosity', 'desire', 'disappointment', 'disapproval', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'grief', 'joy', 'love', 'nervousness', 'optimism', 'pride', 'realization', 'relief', 'remorse', 'sadness', 'surprise', 'neutral']

Metrics (exact match of labels per item)

This is a multi-label, multi-class dataset, so each label is effectively a separate binary classification. Evaluating across all labels per item in the go_emotions test split the metrics are shown below.

Optimising the threshold per label to optimise the F1 metric, the metrics (evaluated on the go_emotions test split) are:

• Precision: 0.378
• Recall: 0.438
• F1: 0.394

Weighted by the relative support of each label in the dataset, this is:

• Precision: 0.424
• Recall: 0.590
• F1: 0.481

Using a fixed threshold of 0.5 to convert the scores to binary predictions for each label, the metrics (evaluated on the go_emotions test split, and unweighted by support) are:

• Precision: 0.568
• Recall: 0.214
• F1: 0.260

Metrics (per-label)

This is a multi-label, multi-class dataset, so each label is effectively a separate binary classification and metrics are better measured per label.

Optimising the threshold per label to optimise the F1 metric, the metrics (evaluated on the go_emotions test split) are:

	f1	precision	recall	support	threshold
admiration	0.540	0.463	0.649	504	0.20
amusement	0.686	0.669	0.705	264	0.25
anger	0.419	0.373	0.480	198	0.15
annoyance	0.276	0.189	0.512	320	0.10
approval	0.299	0.260	0.350	351	0.15
caring	0.303	0.219	0.489	135	0.10
confusion	0.284	0.269	0.301	153	0.15
curiosity	0.365	0.310	0.444	284	0.15
desire	0.274	0.237	0.325	83	0.15
disappointment	0.188	0.292	0.139	151	0.20
disapproval	0.305	0.257	0.375	267	0.15
disgust	0.450	0.462	0.439	123	0.20
embarrassment	0.348	0.375	0.324	37	0.30
excitement	0.313	0.306	0.320	103	0.20
fear	0.550	0.505	0.603	78	0.25
gratitude	0.776	0.774	0.778	352	0.30
grief	0.353	0.273	0.500	6	0.70
joy	0.370	0.361	0.379	161	0.20
love	0.626	0.717	0.555	238	0.35
nervousness	0.308	0.276	0.348	23	0.55
optimism	0.436	0.432	0.441	186	0.20
pride	0.444	0.545	0.375	16	0.60
realization	0.171	0.146	0.207	145	0.10
relief	0.133	0.250	0.091	11	0.60
remorse	0.468	0.426	0.518	56	0.30
sadness	0.413	0.409	0.417	156	0.20
surprise	0.314	0.303	0.326	141	0.15
neutral	0.622	0.482	0.879	1787	0.25

The thesholds are stored in thresholds.json.

Use with ONNXRuntime

The input to the model is called logits, and there is one output per label. Each output produces a 2d array, with 1 row per input row, and each row having 2 columns - the first being a proba output for the negative case, and the second being a proba output for the positive case.

# Assuming you have embeddings from all-MiniLM-L12-v2 for the input sentences
# E.g. produced from sentence-transformers such as:
#      huggingface.co/sentence-transformers/all-MiniLM-L12-v2
#      or from an ONNX version E.g. huggingface.co/Xenova/all-MiniLM-L12-v2

print(embeddings.shape)  # E.g. a batch of 1 sentence
> (1, 384)

import onnxruntime as ort

sess = ort.InferenceSession("path_to_model_dot_onnx", providers=['CPUExecutionProvider'])

outputs = [o.name for o in sess.get_outputs()]  # list of labels, in the order of the outputs
preds_onnx = sess.run(_outputs, {'logits': embeddings})
# preds_onnx is a list with 28 entries, one per label,
# each with a numpy array of shape (1, 2) given the input was a batch of 1

print(outputs[0])
> surprise
print(preds_onnx[0])
> array([[0.97136074, 0.02863926]], dtype=float32)

# load thresholds.json and use that (per label) to convert the positive case score to a binary prediction 

Commentary on the dataset

Some labels (E.g. gratitude) when considered independently perform very strongly, whilst others (E.g. relief) perform very poorly.

This is a challenging dataset. Labels such as relief do have much fewer examples in the training data (less than 100 out of the 40k+, and only 11 in the test split).

But there is also some ambiguity and/or labelling errors visible in the training data of go_emotions that is suspected to constrain the performance. Data cleaning on the dataset to reduce some of the mistakes, ambiguity, conflicts and duplication in the labelling would produce a higher performing model.